Sincronia: Near-Optimal Network Design for Coflows Shijin - - PowerPoint PPT Presentation

Sincronia:

Near-Optimal Network Design for Coflows

Saksham Agarwal Akshay Narayan Rachit Agarwal David Shmoys Amin Vahdat

Shijin Rajakrishnan

Joint work with

The Flow Abstraction

Optimized for Flow-level performance FTP Email HTTP … Traditional Applications: Care about performance

• f individual flows

Good Match

Is Flow Still the Right Abstraction?

Distributed Applications: Care about performance for a group of flows … FTP Email HTTP … Traditional Applications: Care about performance

• f individual flows

Optimized for Flow-level performance

The Coflow abstraction

Collection of semantically related flows [Chowdhury & Stoica, 2012]

… … … … Coflow 1 Coflow 2 Coflow 3

Allows applications to more precisely express their performance goals

• Big-switch model
• Clairvoyant scheduler

▪

Coflow details known at arrival time:

➢ Source-destination for each flow ➢ Size of each flow ➢ Coflow weight

• Metric – coflow completion time: Time when all flows complete

Network and Coflow Model

Ingress ports Egress ports

1 2’ 2 1’

DC Fabric

Goal: Minimize Average Weighted Coflow Completion Time (CCT)

Prior Results

Practical, Near-Optimal Network Design for Coflows? Impossibility Results

• NP-hard
• <2x approximation hard

Systems/ Theory State-of-the-art Performance Guarantees Runs on Existing Transport Work Conserving Starvation Avoiding Systems Varys [SIGCOMM ‘14] Theory On Scheduling Coflows [IPCO ‘17]

(4-apx)

Sincronia:

Given a set of coflows and a “right” ordering, ANY per-flow rate allocation mechanism that is work-conserving & order-preserving produces average CCT within 4x of optimal Guarantees 4-approximation for (weighted) average CCT

• Per-flow rate allocation irrelevant
• Transport layer agnostic

Two key results

Systems/ Theory Name Performance Guarantees Runs on Existing Transport Work Conserving Starvation Avoiding Systems Varys Theory On Scheduling Coflows Systems Sincronia

(4-apx) (4-apx)

Sincronia – Near-Optimal Network Design

Also outperforms state-of-the-art across evaluated workloads

• Algorithm – BSSI

▪ Bottleneck, Select, Scale, Iterate ▪ SRPT-first style algorithm

Sincronia Design

• Priorities set from order
• Flows offloaded to transport layer
• No explicit per-flow rate allocation

Set of coflows Ordered set of coflows Priorities

• n flows

Coflow

• rdering

Flow Scheduling

Bottleneck-Select-Scale-Iterate (BSSI)

• Find BOTTLENECK port
• SELECT (weighted) largest job

▪ Ordered last

• SCALE weights of remaining jobs
• ITERATE on unscheduled jobs

1 2 1’ 2’

Ordering not important

BSSI in Action

• Bottleneck
• Select

▪ Ordered Last

• Scale
• Iterate

1 2’ 2 1’ 1 2 1’ 2’

Order: Weights: Find port handling largest number of packets Select coflow with largest size-to-weight ratio Scale weight of each coflow (at bottleneck port) Iterate on unscheduled coflows Weight ← Weight(1 –

ൗ

Size Weight

ൗ

Size Weight)

Weight ← Weight(1 –

ൗ

Size Weight

ൗ

Size Weight)

ൗ

Size Weight = 3

ൗ

Size Weight = 4

ൗ

Size Weight = 1

ൗ

Size Weight = 8

ൗ

Size Weight = 4

1 3 4 Weight ← × (1 – )

#packets = 4 #packets = 8 #packets = 5 #packets = 7

End-to-End Design(Offline)

1 2’ 2 1’ Order:

Host 1 Host 2 Transport Transport

• Each host knows ordering
• Flows get priority of coflow
• Offloads to priority enabled transport layer

BSSI

Per-flow Rate Allocation is Irrelevant

• Intuition: Sharing bandwidth does not help CCT
• Order-preserving schedule:

Flow blocked iff ingress or egress port serving higher-ordered flow

Given the BSSI ordering, ANY per-flow rate allocation mechanism that is work conserving & order-preserving produces average CCT within 4x of optimal

Avoiding per-flow rate allocation: Implications

• Implement on top of any transport layer

▪ E.g. pFabric, pHost, TCP

• Design and implementation independent of

▪ Network Topology ▪ Location of Congestion ▪ Paths of Coflows

• More scalable

▪ No reallocations upon coflow arrivals/departures

Details in paper

Handling Arbitrary Arrival Times

1 2 4 8 1 2 4 8

• Framework: Khuller, Li, Sturmfels, Sun, Venkat, ‘18
• Time divided into epochs
• In each epoch

▪ Choose subset of unscheduled jobs ▪ Schedule in next epoch using offline alg.

Provides 12-competitive performance (details in paper)

Evaluation Overview

• Testbed implementation on top of TCP

▪ Evaluate impact of in-network congestion, and hardware constraints

• Simulations

▪ Coflows arrive at time 0 ▪ Coflows arrive at arbitrary times

▪ Sensitivity analysis

➢Coflow sizes, structure, # of coflows ➢Network topologies, Oversubscription ratios, Network load ➢…

All simulations, workloads, and implementations are open- sourced on Sincronia website

Simulation Results Offl fline

1 2 3 4 5 6 7 8 9

Average 90th percentile 99th percentile Facebook trace 1000 coflow trace 2000 coflow trace

OCT: Completion time of a coflow in an unloaded network

Sincronia not only provides near-optimal guarantees, but also improves upon state-of-the-art design in practice

526 coflow trace [Varys]

Simulation Results Online

0.5 1 1.5 2 2.5 3 3.5 4

1000 coflow trace 2000 coflow trace

Average 90th percentile 99th percentile

Slowdown

Network Load = 0.9 Even at such high network loads, Sincronia achieves CCT close to that of an unloaded network

Implementation Results

Implemented on top of TCP

• 16-server Fat tree topology

▪ Full bisection bandwidth ▪ 20 PICA8 switches

➢ Supports 8 priority levels

• DiffServ for priority scheduling

Implementation Results

20 40 60 80 100 120 140 160

• Unfair Evaluation
• TCP not designed for coflows
• TCP not designed to minimize CT

+ Compare against existing designs

• E.g. Varys reports 1.85x improvement

at mean and at tails

Average 90th percentile 99th percentile

Sincronia achieves significant improvements over existing network designs even with a small number of priority levels

Summary

• Sincronia – a network design for coflows
• 4x within optimal
• No per-flow rate allocation

Name Performance Guarantees Run on existing Transport Work Conserving Starvation Avoiding Varys On Scheduling Coflows Sincronia

(4-apx) (4-apx)

• Paper discusses number of open problems

Thanks!